For decades, a network was a set of boxes, each configured by hand through its own command line. Today most traffic crosses networks that are defined, deployed and changed by software: cloud VPCs created by an API call, virtual firewalls spun up in seconds, Kubernetes clusters wiring thousands of Pods automatically. This lesson connects the classic concepts from earlier lessons to those modern forms.
Three planes
Every network device does three kinds of work:
| Plane | Job | Example |
|---|---|---|
| Data plane (forwarding) | Move each packet out the right port, fast | Switching ASIC, Linux kernel forwarding |
| Control plane | Decide the paths and build the tables | OSPF, BGP, spanning tree |
| Management plane | Configure and monitor the device | SSH CLI, SNMP, APIs |
Traditionally all three live inside every box, and each box computes its own view of the network.
Software-defined networking (SDN)
SDN separates the control plane from the devices and centralizes it in a controller that sees the whole network. Switches become simple forwarding engines programmed by the controller.
- Southbound interfaces let the controller program devices: OpenFlow (the original SDN protocol, with match-action flow tables), NETCONF, gNMI, or vendor APIs.
- Northbound APIs let applications and operators ask for outcomes: “connect these two apps”, “isolate this tenant”.
Benefits: network-wide visibility, consistent policy, quick changes through an API, and less vendor lock-in. Where it lives today:
- Cloud networks: every VPC is SDN; you never configure a physical switch at AWS, Azure or Google.
- Data center fabrics: controllers or EVPN-based automation (Cisco ACI, Arista CloudVision, open-source SONiC).
- SD-WAN: a controller picks the best path (fiber, broadband, 5G) per application across branch offices.
- Kubernetes networking: CNI plugins such as Cilium and Calico program the forwarding of every node from cluster state.
$ sudo ovs-vsctl show
6e3c9a4e-…
Bridge br-int
Port vm1-tap
Interface vm1-tap
Port br-int
Interface br-int
type: internal
$ sudo ovs-ofctl dump-flows br-int | head -2
cookie=0x0, table=0, n_packets=1842, priority=100,in_port="vm1-tap",dl_src=52:54:00:7a:11:02 actions=resubmit(,10)
cookie=0x0, table=0, n_packets=0, priority=0 actions=drop
Open vSwitch is a programmable software switch widely used in virtualization and OpenStack; dump-flows shows the OpenFlow match-action rules a controller installed.
Network function virtualization (NFV)
NFV replaces dedicated appliances (firewalls, load balancers, routers, VPN gateways, mobile core functions) with software running on ordinary servers, as VMs or containers: virtual network functions (VNFs) or, when containerized, cloud-native network functions (CNFs).
| Physical appliance | VNF / CNF | |
|---|---|---|
| Deploy time | Weeks (order, ship, rack) | Minutes |
| Scale | Buy a bigger box | Run more instances |
| Cost | Specialized hardware | Commodity servers |
| Performance | Hardware-accelerated | Needs tuning (SR-IOV, DPDK, eBPF/XDP) |
SDN and NFV complement each other: NFV creates the functions, SDN steers traffic through them (service chaining: firewall, then IDS, then load balancer). Telecom operators run their 5G cores this way.
Overlays: VXLAN
Cloud and data center networks give each tenant its own isolated layer 2 or layer 3 network on top of a shared physical one. VXLAN does this by wrapping Ethernet frames in UDP (port 4789) with a 24-bit VNI, allowing about 16 million segments instead of 4,094 VLANs. The physical network (the underlay) only routes IP between hosts; the overlay networks live in the encapsulation. EVPN (a BGP extension) often acts as the control plane that tells each host where every MAC address lives.
$ sudo ip link add vxlan100 type vxlan id 100 dstport 4789 local 10.0.0.11 remote 10.0.0.12 dev eth0
$ sudo ip link set vxlan100 up
$ ip -d link show vxlan100 | grep -o 'vxlan id [0-9]* remote [0-9.]*'
vxlan id 100 remote 10.0.0.12
Remember the overhead: 50 bytes of extra headers, so overlays need an underlay MTU of at least 1,550 or an overlay MTU of 1,450.
Cloud service models
| Model | You manage | Provider manages | Examples |
|---|---|---|---|
| IaaS (infrastructure) | OS, runtime, apps, data, network config | Hardware, virtualization, physical network | EC2, Azure VMs, Compute Engine |
| PaaS (platform) | Apps and data | Everything below the app | App Engine, Heroku, managed Kubernetes and databases |
| SaaS (software) | Your data and settings | The whole application | Microsoft 365, Gmail, Salesforce |
| FaaS (functions, serverless) | Function code | Servers, scaling | AWS Lambda, Cloud Functions |
Deployment models: public cloud, private cloud, hybrid (both, connected) and multi-cloud (several providers).
Cloud networking in one picture
Every concept from this course reappears in a VPC (virtual private cloud):
| Cloud building block | Classic concept |
|---|---|
VPC with a CIDR such as 10.40.0.0/16 | Your private address plan (lessons 7 and 8) |
| Public and private subnets per availability zone | Subnetting and VLANs |
| Route tables, internet gateway | Routing tables and default routes (lesson 11) |
| NAT gateway | PAT for private subnets (lesson 13) |
| Security groups and network ACLs | Stateful and stateless firewalls (lesson 21) |
| Load balancers (L4 and L7) | Load balancers and reverse proxies (lessons 2 and 18) |
| Peering, transit gateways, VPN, Direct Connect / ExpressRoute | Routing between networks, IPsec, private WAN circuits |
| Private DNS zones | Split-horizon DNS (lesson 17) |
$ aws ec2 describe-route-tables --route-table-ids rtb-0a1b2c3d4e5f67890 \
--query 'RouteTables[0].Routes[].[DestinationCidrBlock,GatewayId||NatGatewayId]' --output text
10.40.0.0/16 local
0.0.0.0/0 nat-0f1e2d3c4b5a69788
That private subnet routes the VPC locally and everything else to a NAT gateway: exactly the home router pattern, defined by API.
IoT networking
The Internet of Things connects sensors, meters, cameras, machines and vehicles. Its constraints differ from laptops: tiny CPUs, batteries that must last years, lossy radio links and huge device counts.
| Need | Typical choice |
|---|---|
| Short range, home and building | BLE, Zigbee, Thread, Wi-Fi HaLow |
| Long range, very low power, little data | LPWAN: LoRaWAN, Sigfox, NB-IoT, LTE-M |
| Messaging | MQTT (lesson 19), CoAP (a compact REST-like protocol over UDP) |
| Addressing | IPv6, often with 6LoWPAN compression on 802.15.4 |
| Gateways | Translate between local radios and IP, buffer data, run local logic (edge computing) |
Security is the hard part: default passwords, unpatched firmware and devices that will run for a decade. Put IoT devices on their own VLAN or SSID, allow them only the destinations they need, and change every default credential.
5G network slicing
Network slicing runs several logical networks on one physical 5G infrastructure, each with its own guarantees, built from SDN and NFV end to end (radio, transport and core):
| Slice | Optimized for | Example |
|---|---|---|
| eMBB | Throughput | Phones, fixed wireless access |
| URLLC | Latency and reliability | Factory robots, remote surgery trials |
| mMTC | Huge numbers of low-rate devices | Smart meters, asset trackers |
A broadcaster might buy a slice with guaranteed uplink capacity for live video at a stadium, isolated from the crowd’s phones in the same cells. Slicing needs a standalone 5G core.
Network automation
Hand-typed configuration does not scale and drifts over time. Modern teams manage networks like software:
- Source of truth: an inventory and IP plan (NetBox, Nautobot) that says what should exist.
- Infrastructure as code: Terraform for cloud networks, Ansible or Nornir for devices, configs reviewed in Git.
- Model-driven APIs: NETCONF and RESTCONF with YANG models, and gNMI for streaming telemetry.
- Validation: tools like Batfish test configurations before they are deployed, and CI pipelines run those tests on every change.
- name: Ensure VLANs exist on access switches
hosts: access_switches
gather_facts: false
tasks:
- name: Configure VLANs
cisco.ios.ios_vlans:
config:
- { vlan_id: 10, name: STAFF }
- { vlan_id: 20, name: GUEST }
state: merged
The Terraform Tutorial and the Kubernetes Tutorial on this blog show these ideas in practice.
Exercise
What does SDN separate from the network devices?
Show answer
Answer: C. In SDN, devices keep fast forwarding (data plane) while a controller with a network-wide view computes paths and programs them.
Exercise
A company replaces its hardware firewall appliances with firewall software running in VMs on standard servers. What is that called?
Show answer
Answer: A. NFV turns network functions into software (VNFs or CNFs) on commodity hardware, so they can be deployed and scaled like applications.
Exercise
Why do cloud providers use VXLAN or similar overlays instead of plain VLANs for tenant networks?
Show answer
Answer: D. Multi-tenant clouds need millions of isolated networks. VXLAN carries them over a routed IP underlay. Encryption is a separate layer.
Key takeaways
- Networks have data, control and management planes; SDN centralizes control in a controller.
- NFV turns appliances into software; SDN steers traffic through them.
- Overlays like VXLAN build millions of tenant networks over a routed underlay.
- A cloud VPC is subnets, route tables, NAT, firewalls and load balancers, defined by API.
- IoT needs low-power radios and strict segmentation; 5G slicing delivers isolated logical networks; automation treats network config as code.
FAQ
Do I need to learn SDN to work with the cloud?
You use it every time you create a VPC or security group, without seeing the controller. What matters is the classic concepts underneath (subnets, routing, NAT, firewalls), which this course covers.
Is OpenFlow still used?
Less than its early hype suggested. The SDN ideas won, but many deployments use other southbound protocols, EVPN, or vendor controllers. Open vSwitch and research networks still use OpenFlow widely.
Is edge computing the same as a CDN?
Related but broader. A CDN caches content near users; edge computing runs application code near users or devices, for example processing camera video at a factory instead of sending it to a distant cloud region.
Where should a network engineer start with automation?
Start by putting configs in Git, then automate read-only tasks (backups, inventory, audits) with Ansible or Python, and only then automate changes, with review and testing in a pipeline.